The Problem: AI Systems That Ignore the Rules
Our client is a large European insurance company. They had an expensive problem: their claims systems made decisions that did not match their policy documents.
They manage over 3,000 different policies for cars, homes, and businesses. That is more than 50,000 pages of legal text. Each policy has its own rules, exceptions, and limits. Even experienced employees found it hard to apply these rules correctly every time.
A customer had an accident and filed a claim. Their policy (AUTO-PREMIUM-2024) covers personal use only. The AI system approved the payment.
The problem: The customer was working as a delivery driver at that moment. Section 4.2.1 of the policy says: no coverage for commercial delivery work. But the system did not check this rule. The company paid €8,400 that they should not have paid.
This was not a single mistake. Similar wrong decisions happened hundreds of times per month.
The compliance team was worried. Every wrong decision could mean problems with regulators. Every wrong denial could lead to a lawsuit. And when auditors asked “why was this claim denied?” the answer was often “the AI decided.” Regulators do not accept this answer.
The Solution: LLM + Knowledge Graph
The client had tried two approaches before. Rule-based systems were too strict and could not understand normal language. Pure LLM solutions were too creative and invented rules that did not exist. They needed both: the flexibility of language models and the precision of formal rules.
This is what Synapse OS does. It is a hybrid system. The LLM understands human language. But every answer is checked against a knowledge graph that contains the real rules. The LLM suggests. The graph confirms.
Phase 1: Reading the Documents with LLM
We used large language models to read all 3,000+ policy documents. But the goal was not to summarize them. The LLM's job was structured extraction. For example: finding that “Section 4.2.1 Exclusions” in AUTO-PREMIUM-2024 contains 14 specific conditions. Each condition becomes a formal rule in the graph. The system also maps how rules connect to each other.
Lawyers reviewed what the LLM extracted. About 3% of the rules needed corrections. They fixed these errors once. After that, the corrected rules became permanent company knowledge. This is not training data that can change. These are fixed rules in a graph database.
Phase 2: Making Decisions at Runtime
When a new claim arrives, the system works in three steps:
- LLM understands the claim — A customer writes “I had an accident while dropping off a package for my side gig.” The LLM understands that “side gig” means commercial delivery work.
- Graph checks the rules — The system asks the knowledge graph: “Does AUTO-PREMIUM-2024 cover accidents during commercial delivery?” The graph returns NO, with the exact rule and section number.
- LLM writes the response — The denial letter is written in normal language. But every statement is connected to a rule in the graph that points to the original document.
The LLM cannot invent rules. If a rule is not in the graph, the system cannot use it. No graph reference means no output.
RULE: Section 4.2.1, Paragraph 3
SOURCE: AUTO-PREMIUM-2024.pdf, Page 127, Lines 14-18
The LLM handles the language part: understanding that “side gig” means commercial work. The graph handles the rules part: knowing that commercial use means exclusion 4.2.1.3 applies. Each part does what it does best.
The Key Feature: Deep Root Cause Analysis
The compliance team was most excited about the depth of explanation. Real root cause analysis (RCA) is not just “here is the rule.” It is a chain of “why” questions that traces a decision back to its origin.
When an auditor or customer questions a denial, they want to understand the complete reasoning. Not just one answer. Synapse OS gives them the full chain.
This is more than documentation. It is business intelligence. The analysis does not just explain why a claim was denied. It traces the problem back to a gap in the sales process. This gives the business useful information to prevent future problems.
“Why was the claim denied?”
“Commercial use is not covered.”
End of explanation. Customer complains. Legal team gets involved. 2 weeks of emails and calls.
“Why was the claim denied?”
5-level trace. Every step explained. Every rule has a source. Every data point is documented.
Customer understands. Problem solved in one phone call. Feedback sent to improve sales process.
Results After Six Months
100% Regulatory Compliance
Zero compliance violations since the system went live. When regulators did their quarterly audit, the response time changed from one week of work to same-day delivery. The Chief Compliance Officer said: “This is the first audit in five years where I could sleep well.”
42% Fewer Processing Errors
Claims that should have been denied (like the delivery driver case) are now caught automatically. Claims that should have been approved now get approved with clear reasons. Result: fewer reversals, fewer disputes, fewer lawsuits.
Audit Time: Days → Seconds
Before, the compliance team spent 3-4 days preparing for each audit. They had to manually trace decisions back to policy documents. Now they click one button. The complete decision trace appears in less than one second.
“For the first time, we can tell a regulator exactly why a claim was denied. We show the exact paragraph and page number. We do not explain what the AI ‘thought.’ We show them the rule in the document.”
Why This Worked: The Hybrid Approach
The insurance industry tried two approaches before. Both failed:
Very precise, but very limited. Could not understand “I was doing my side hustle” because nobody programmed that exact phrase. Required constant manual updates.
Very flexible, but not reliable. Sometimes the LLM would cite “Section 7.3” that does not exist. Or it would approve claims based on feelings instead of policy rules.
Synapse OS combines both: LLMs for understanding language, knowledge graphs for correct rules.
The LLM understands different ways people say things. “Side gig,” “freelance delivery,” and “Uber Eats job” all mean the same thing. But the LLM does not make the final decision. It asks the knowledge graph. The graph contains verified rules that lawyers have approved, extracted from the actual policy documents.
The result: A system that is as flexible as ChatGPT, but as precise as a legal database. It understands what you mean. It knows what the policy says. And it can prove the connection between the two.